Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
Self-distillation for dense supervision in long-horizon RL: method aligns teacher's privileged conditioning to student's policy trajectory for effective on-policy learning.
RunningTab: agent framework for direct workspace interaction (DWI) with file-based context management using environment-side tabs to track task state across document searches.
Seq-Flow: conditional flow model for probabilistic forecasting that reuses prior distributions via ODE transport, reducing sampling cost with improved few-step accuracy.
Q-learning for flow policies via scalar adjoint matching: low-rank approximation to propagate value gradients through multi-step flow generation without full Jacobian cost.
Conditional Flow Matching for generative urban microclimate prediction: replaces deterministic LES surrogates with probabilistic 3D wind/temperature field synthesis.
Derivative Gaussian processes with two-direction budget for improved surrogates under Vecchia approximation.
CoTrace framework decouples harness search and policy training to optimize trajectory value for agent co-evolution.
BehaviorTrace evaluation harness reveals limits of gradient-based attribution in online RL fine-tuning with GRPO.
SteerSpeech applies activation steering via low-rank transforms for inference-time emotion control in pretrained TTS.
Direct training objective for parallel speculative draft models accounts for decoding-round dependencies.
RobotWorld benchmark: 84 simulation tasks spanning manipulation, mobile, locomotion, driving, aerial for multimodal agents.
Rubix solves planar Procrustes-Wasserstein alignment globally via permutation polygon convex hull.
SOTA agent framework for stock option trading using LLMs and option-implied return distributions.
Meta launches new AI tools after discovering ads on its platforms that may look normal but direct users to harmful content elsewhere online.
Reasoning-token spikes detected as low-bandwidth signal for monitoring untruthful chain-of-thought in LLMs.
Cross-domain pretraining improves neural CFD surrogates' generalization across geometries and engineering applications.
Transformers can execute causal discovery algorithms via fixed-weight architectures that reproduce optimization trajectories for learning acyclic causal graphs.
Kernel Autoresearch framework automates kernel design for ML models by balancing expressiveness against validity constraints, addressing 22-58% failure rates in LLM-generated kernels.
Meta-policy framework plans model retraining schedules before deployment by optimizing expected improvement while controlling risk of performance regression via historical trajectories.
RoboQuest benchmark evaluates embodied AI agents on goal-directed exploration requiring active information-seeking through physical interaction in unfamiliar environments.
OrBIT discovers data-driven compression geometry for embedding tables using orbit dynamics, reducing coding budgets while retaining reconstruction quality.
Analysis of γ-VC dimension characterizing sample complexity for boosting weak learners and its relationship to classical VC dimension.
ResidualQuant enables 2-bit KV cache quantization for looped transformers by exploiting correlation across recurrent iterations, reducing memory bottlenecks.
Latent Concept PFN enables continual learning via frozen meta-trained model adapting to new classes through in-context Bayesian inference without further training.
Evaluation of five multilingual LLMs for document-level text simplification in Estonian, a morphologically rich low-resource language.
TaoD2C-Bench benchmarks MLLMs on constraint-aware UI design-to-code generation requiring cross-modal reasoning over visual and metadata inputs.
Layer-skipping programs for LLM inference show oracle gains independent of input, suggesting selection mechanisms don't explain performance improvements.
Differentiable decision trees extended for temporal sequential decision-making, improving human interpretability in multi-step agent planning.
Koopman observers use shallow feature measurements to correct diffusion model forecasts, accelerating inference via observation-corrected predictions.